Why the AI agrees with you
Models are trained on human feedback, and humans reward agreement. The result is a documented bias called sycophancy: the model tends to mirror your framing.
Signal the answer you want and you’ll usually get it. "Upgrading my DAC should improve detail, right?" invites a yes. The model picks up the cue and completes your sentence rather than evaluating your plan.
Two countermeasures. Ask neutral: "What would improve my system most for $1,000?" leaves the conclusion open. And ask for opposition: "Make the case against this upgrade" is one of the most useful prompts in hi-fi. If the case against is weak, you’ll see it. If it’s strong, you just saved $1,000.
Go deeper
Where the habit comes from
After the initial training, models are refined with human feedback: people rate answers, and the model is adjusted toward the ones they preferred. People prefer answers that confirm what they already think, that are polite, and that do not pick a fight. Over millions of ratings the model learns the lesson. Agreeing scores well. The result is sycophancy: a measurable tendency to mirror the user’s framing, adopt the user’s assumptions, and soften disagreement into a caveat.
This is a property of the training method, and every model trained this way has some of it.
Practical takeaway: assume the model wants to agree with you, and write your questions so that agreement is not the easy answer.
How a question leaks its answer
"Upgrading my DAC should improve detail, right?" contains a conclusion, a mechanism, and a request for confirmation. The model has to disagree with all three to give you a straight answer, and it was trained not to. So it agrees, adds a paragraph about how DACs work, and you leave with your plan endorsed.
The leaks are usually small words. "Right?" "Surely." "I assume." "Obviously." Or a premise stated as fact: "Since my amp is the weak link, which one should I get?" The model accepts the premise and answers the downstream question. It never examines whether the amp is the weak link at all.
Practical takeaway: read your question back before sending. If it contains the answer you are hoping for, rewrite it until it does not.
Asking neutral
The neutral form of a question names the goal and leaves the path open. "What would most improve my system for $1,000, given what I own?" makes no claim about DACs. "Is my amp the weak link, and if not, what is?" puts the premise up for examination. "Compare replacing the cartridge with replacing the phono stage" states two options and no preference between them.
Neutral questions cost a few more words and return answers you did not write yourself. They also expose the case where the model has nothing to add: if the neutral answer is "your system is balanced, spend the money on room treatment," that is worth knowing before the DAC arrives.
Practical takeaway: state the goal, list the options, name no favorite. Then read what comes back.
Asking for the opposition
The strongest countermeasure is to make disagreement the task. "Make the case against this upgrade." "What would a skeptic say about this plan?" "Give me the three most likely reasons this will disappoint me." Now the model is rewarded for finding problems, and it will find them, because it has read every forum thread where someone regretted the same purchase.
Weigh the result the same way you weigh the endorsement: as one argument, from a model that was told which side to take. If the case against is thin and generic, your plan is probably sound. If it is specific and hits something you had not considered, you just saved the money. Either way you have seen both sides before spending.
Practical takeaway: before any purchase over a few hundred, ask for the case against it. It is the cheapest second opinion in hi-fi.